Generative Adversarial Neural Networks Model of Photonic Crystal Fiber Based Surface Plasmon Resonance Sensor

Instantaneous Property Prediction and Inverse Design of Plasmonic Nanostructures Using Machine Learning: Current Applications and Future Directions

“…Unsupervised pre-training also contributes to higher quality networks provided with small data sets [61]. Semi-supervised training methods such as generative adversarial networks are more widely used and proven useful in reducing required training data [66,67].…”

“…With the invention of generative adversarial networks (GANs) [78], there have been increasing cases of using GANs individually or in parallel with ANNs in designing photonic and plasmonic devices [66,77]. Interestingly, GAN consists of two networks-the discriminator and the generator playing the roles of the detector and the counterfeiter (Figure 3d).…”

“…The detector, as the name suggests, is trained to distinguish the fake generated data from the pool of data input, and the counterfeiter is trained with the feedback from the detector to counterfeit the product as close as possible to the ground truth. In recent research, GAN can not merely generate proper nano-photonic structures designed for certain functionalities [56] but also generates additional training data when the available dataset is insufficient [66].…”

“…Data gathering in developing ML algorithms for plasmonics is an inevitably long process, as researchers mostly use the traditional simulation software to provide the "ground truth" for the network to refer to in supervised learning models [54,77,84,101]. To improve the efficiency of data gathering and the training quality, researchers have augmented ANN with GAN, which generates counterfeit data mimicking the real data to expand the training dataset [66,78]. With the generative nature of GAN, the training error of ANN is relatively lower and more stable compared to that without being connected to GAN (Figure 6e).…”

“…(e) Performance comparison between GAN-augmented ANN and regular ANN with a lean dataset.The training Mean Squared Error of GAN-augmented ANN is comparatively small and stable. Reproduced with permission from Ref [66]…”